Open set recognition for drone based on deep metric learning

Abstract

With the rapid development of drone technology and its wide range of applications, some potential security issues have gradually surfaced. In practice, however, most drone identification methods are based on the closed set assumption, which hinders the capacity to respond effectively to unidentified drone classes.To address this issue, we propose an open set recognition method based on deep metric learning in this paper. Specifically, we design a multi-scale time-frequency feature extraction network that learns and extracts low-dimensional embedding representations from the Time-Frequency Spectrum (TFS) of Unmanned Aerial Vehicle (UAV) radio-frequency signals. In the subsequent feature optimization stage, we jointly employ triplet loss and center loss to refine the feature distribution. Moreover, we directly use the distances between the optimized class centers and the feature vectors as the basis for classification, leading to a simple and deployment-friendly prototype-distance based recognition rule that avoids introducing an additional trainable classifier module and does not require test-time iterative clustering to form prototypes. Finally, we propose an Adaptive Threshold Distance Distribution Classifier (AT-DDC), which dynamically determines the optimal decision threshold by analyzing the distance distribution between feature vectors and class centers, thus enabling open set recognition of drones.Compared with the existing method, the proposed method’s known average accuracy (KA) has been improved by 3.5% in all scenarios. Meanwhile, the true unknown rate (TUR) is reduced by only 1.40%, and the mean absolute percentage error (MAPE) between the decision thresholds and the optimal thresholds is reduced by 8.68%, which demonstrates the excellent performance of the proposed method in drone open set recognition.